Acoustic event detection for auditory context inference

Abstract

Environmental sounds are an ample source of information that could be used to recognize the context of any person in daily life. Usually, for monitoring and surveillance, an Intelligent Transportation System employs various types of infrastructure-based technologies into vehicles and roadways. But these solutions have high installation, maintenance, and operational costs. To overcome these problems, few smartphone sensing based solutions have been proposed, but these are mostly based on the use of positions sensors, motion sensors, or a combination of few sensors; and need a substantial amount of time for detection. In this thesis, we have proposed a cost-effective approach to infer the traffic state of the road and transportation mode of the commuter by analyzing the Cumulative acoustic signal collected from the Microphone sensor of the commuter s smartphone. Different types of sounds are obtained on recording an acoustic signal on a road-segment or inside a vehicle. These sounds have varied spectral contents and provide significant cues to differentiate between various traffic states and transportation modes. In the first part of the thesis, we have presented a detection technique using Wavelet Packet Transform features and Support Vector Machine classifier, combined with crowdsourcing to classify the acquired datasets of traffic noises and vehicle noises. In the second part of the thesis, for improving the detection accuracy and robustness, we have proposed a deep learning based detection approach. For validating the proposed approaches, various field experiments were conducted on the acoustic datasets collected using different smartphones under varied conditions in different cities of India. From the experimental results achieved, it can be concluded that the proposed deep learning based traffic state and transportation mode detection is effective and provides the highest overall accuracy in comparison to the hand-engineered features.